On-patient medical record and mRNA therapeutics using intradermal microneedles

Main

Medical interventions often require specifically timed doses, necessitating accurate medical record-keeping. In particular, mRNA-based delivery systems have proved to be a versatile platform for vaccine and therapeutics development against incurable diseases^1,2,3,4,5, often requiring multiple doses. To name a few examples, clinical trials for RNAactive CV7201 for rabies (CureVac)⁶, RNActive CV9103 for prostate cancer (CureVac)⁷ and mRNA-4157-P201 for melanoma (Moderna)⁸ required 3, 5 and 9 dose administrations, respectively. Many other therapeutics and most of the recommended childhood immunizations require multiple doses (for example, the hepatitis B and polio vaccines) as well^9,10,11. Failure to complete these timed doses risks suboptimal protection. Currently, as many as 40% of patients fail to comply with medical treatments around the globe, and poor adherence is responsible for 125,000 annual deaths in the United States alone^12,13. In sub-Saharan Africa, 35% of children between 12 and 23 months fail to complete recommended childhood vaccinations^14,15. Several factors contribute to these gaps, including lack of access or inability to afford medical treatments and vaccines, but a major contributor is inadequate medical record-keeping systems^16,17. Conventional methods like paper-based cards and online databases present serious risks of losing access to medical histories, and therefore the risk of missing or mistimed follow-up doses. For that reason, a handful of medical record-keeping means have emerged based on fingerprint scanning¹⁸, cell phone applications^19,20,21, microchips²² and more. However, these approaches have raised privacy concerns related to individual, identifiable medical data storage in a centralized database, which poses a risk of data breech, misuse or quality implications^{23,24,25,26,27}.

We developed a robust on-patient medical record-keeping (OPMR) technology using a dissolvable microneedle patch (MNP) that delivers a quantum dot (QD)-based near-infrared (NIR) fluorescent dye encapsulated in poly(methyl methacrylate) (PMMA) microparticles into the skin to encode medical information²⁸ (Fig. 1a). This dye, once deposited into the dermis, is invisible to the naked eye, offering patient data privacy and anonymity, but provides discrete NIR signals that can be detected using a NIR imaging system (Extended Data Fig. 1a–c). By depositing the dye in a predefined pattern that correlates to a specific set of information, the technology can be imaged by healthcare workers to support next-dose decisions without requiring internet connectivity or the use of centralized databases.

**Fig. 1: Schematic of the OPMR technology used for medical information record-keeping.**

Herein, to enable the OPMR with excellent information capacity, security and reliability, we designed the MNP architecture and administration for consistent and optimal data transfer and longevity; achieved an information capacity of billions of encoded patterns using an error correcting code; and developed a temporally and spatially reliable information retrieval system using machine learning. Further, we successfully co-delivered the OPMR with a potent mRNA vaccine encapsulated in lipid nanoparticles (LNPs) that encodes the SARS-CoV-2 spike protein (Fig. 1a). This demonstrates that our OPMR–mRNA MNP technology can co-deliver mRNA therapeutics and corresponding medical information simultaneously. The scope of its application is potentially expandable to any mRNA therapeutics, considering its biocompatibility with mRNA-LNPs and its large encoding capacity in the range of 10⁶ to 10⁹ that accommodates the fast-growing number of mRNA therapeutics under development (Fig. 1b and Extended Data Fig. 1d). This tool could help healthcare workers make informed decisions on follow-up doses in the field where reliable record-keeping is unavailable, and therefore improve medical adherence and complete immunization for the global population.

OPMR MNP materials and architecture

The NIR signal encoding was generated from CuInS₂/ZnS QDs with a photoluminescence quantum yield of 77% and photoluminescence intensity peak at 890–897 nm (ref. ²⁸; Fig. 2a). The QDs were encapsulated in PMMA microparticles to increase the size of the particles and thus mitigate biological clearance and enhance the stability and biocompatibility of the system^28,29,30,31. The PMMA encapsulation did not result in a shift of the peak emission wavelength (Fig. 2a) or major decrease of the photoluminescence quantum yield, as it remained at 73% post-encapsulation. The diameter of the QD–PMMA microparticles (OPMR dye) was tuned to be roughly 10 μm, and the average size was confirmed with scanning electron microscopy (SEM; Fig. 2b). For the evaluation of effective OPMR delivery, MNPs with a 10 × 10 array were used, and were loaded with OPMR dye at the needle tips and a polymer blend as backing (Extended Data Fig. 1e). Each of the 100 microneedles contains OPMR dye at the tip (Fig. 2c), and collectively they form a 10 × 10 array, with each needle corresponding to a single NIR bit (Fig. 2d). Because accurate intradermal information delivery is a critical first step, the MNP application and architecture were designed for consistent dye transfer and optimal signal durability. Towards that end, the three most critical parameters were investigated: (1) bit transfer, the percentage of NIR-labelled needles transferred to and detectable in the skin as bits out of the total number of applied dye-loaded microneedles, as this indicates the initial information encoding to the skin; (2) penetration depth, the intradermal depth where the needle tips deposit the dye, because this may affect signal durability; and (3) needle dissolution, as this translates to the amount of cargo delivered to the skin. For these studies, four needles in one corner of a 10 × 10 MNP were removed for patch orientation, leaving the patch with 96 needles.

**Fig. 2: MNP materials, design and application for effective OPMR delivery.**

First, we compared the results of applying MNPs by hand and with spring applicators (Fig. 2e and Extended Data Fig. 2a). The hand application resulted in poor and inconsistent dye transfer when inspected with the NIR imaging system (Fig. 2f), with less than 20% bit transfer, while applicators exhibited 100% bit transfer (n = 3–5; Fig. 2g and Extended Data Fig. 2b). For penetration depth, needle tips did not reach deeper than 250 μm with the hand application, leaving most of the dye deposited near the epidermis (Extended Data Fig. 2c) and resulting in most of the NIR signals being gone after one month (Extended Data Fig. 2d). This led to an assumption that a deeper dye deposition with a proper applicator is necessary for durable NIR signals. Towards that end, custom spring applicators with a range of impact velocities and holding pressures (Extended Data Fig. 2e) were tested in ex vivo pig skin (Extended Data Fig. 2f), leading to the selection of an impact velocity of 1,407 cm s^–1 and holding pressure of 1.1 MPa for use in the rest of the studies (Extended Data Fig. 2g).

Second, we optimized two MNP design variables that affected the MNP performance: (1) needle tip angle, which correlates to the needle width-to-height ratio and sharpness of the microneedle, and (2) pitch, which is the spacing between two needles from centre to centre (Fig. 2h). Four different tip angles, 8°, 15°, 30° and 53° with a fixed 1.5 mm needle height (Extended Data Fig. 2h), and four different pitches, 0.5 mm, 1 mm, 1.5 mm and 3 mm with a fixed 15° tip angle (Extended Data Fig. 2i), were tested for the maximum bit transfer, penetration depth and needle dissolution while maintaining the minimum patch size. For all four tip angles, penetration depths were well within the dermis layer³², but only the 15° and 30° tip angles resulted in 100% bit transfer (n = 4–7; Fig. 2i). The 8° tip angle was too fragile (Extended Data Fig. 2j) whereas the 53° tip angle was too blunt upon penetration. For needle spacings, all pitches deposited the dye within the dermis, but only the MNPs with a pitch equal to or larger than 1 mm resulted in 100% bit transfer (n = 4–7; Fig. 2j). For needle dissolution, a decreasing trend was observed with increasing tip angles, and an increasing trend was observed with increasing pitches (n = 4–7; Fig. 2k).

OPMR MNP applications for effective delivery

Once the dye particles are deposited in the skin, the NIR signals must last to allow accurate information readings. For this purpose, the tip angles (15° and 30°) and pitches (1 mm and 3 mm) that resulted in 100% bit transfer were tested for signal durability in vivo. Signal durability involves two parameters: (1) signal retention, the percentage of detectable NIR bits out of the total transferred bits, and (2) signal intensity, the pixel brightness value (in a.u.) of NIR bits. Signal retention and signal intensity are key quantitative and qualitative indications of signal durability, respectively. For this study, a Yorkshire pig model was chosen due to its similar epidermis–dermis skin structure and mechanical properties compared to human skin³³ (Extended Data Fig. 2k). Adaptive threshold algorithms were used for the signal retention (Extended Data Fig. 3a) and signal intensity (Extended Data Fig. 3b) analyses³⁴. When three MNP groups (15°–1 mm, 15°–3 mm and 30°–1 mm) were applied (Extended Data Fig. 3c) and imaged weekly for 10 weeks (Extended Data Fig. 3d), the 15° groups resulted in 90.5 ± 9.04% signal retention on week 10, whereas the 30° group exhibited an inferior signal retention of 75.67 ± 22.12% on week 9 (n = 4–6; Fig. 2l). Considering that the 15°–1 mm and 15°–3 mm groups penetrated the skin more deeply (640 μm and 1,350 μm, respectively) than the 30°–1 mm group (300 μm; Fig. 2i,j), penetration depth seems to affect intradermal signal retention. When signal intensities of the two best performing groups (15°–1 mm and 15°–3 mm) were analysed for 70 days, although the overall intensity gradually decreased over time, no substantial difference occurred between the two groups (n = 3–5; Fig. 2m), indicating that signal intensity is not affected as long as dye is deposited deeper than a threshold depth (for example, 600 μm). Because 1-mm-pitch MNPs performed just as well as 3-mm-pitch MNPs while offering a nine times smaller patch size, a 10 × 10 MNP design with 1 mm pitch and 15° tip angle (0.4 mm × 0.4 mm × 1.5 mm) was selected for the rest of the studies. As a result, these optimized application and architecture parameters yielded a consistent penetration depth near 700 μm (Fig. 2n), effectively depositing the dye particles well within the dermis³² (Fig. 2o).

Error correcting code for temporally robust encoding

Our OPMR technology encodes information by imprinting two-dimensional (2D) patterns on MNPs, leveraging the binary feature of microneedle bits (that is, a microneedle (bit) is either present (ON) or absent (OFF)). The OPMR dye, once deposited in the skin, may experience NIR signal reduction due to phagocytic clearance, photobleaching or physical damage such as injury or scarring. To mitigate these potential data corruptions, our system features (1) an error correction scheme that introduces redundancy to compensate for temporal signal decay and (2) deep learning-based image processing to ensure reliable pattern readings despite spatial variations. The overall pipeline consists of two phases: an encoding phase and a decoding phase (Extended Data Fig. 4a,b).

The encoding phase was developed to mitigate potential MNP bit losses that may cause false representation of a pattern over time, and thereby provide temporal robustness. During encoding, information of interest gets converted to a pattern that can be encoded on a MNP. First, the information to be recorded on the patient was determined (Fig. 3a). Then, it was translated to an encoded binary string (Fig. 3b) with an error correcting code (ECC). The ECC added redundancy to the information bits as a means to defend against data corruption and therefore ensure reliable long-term information recovery³⁵. Among the many types of ECCs, a Reed–Muller (RM) code^36,37 (Extended Data Fig. 4c) was selected for the OPMR due to its suitability for correcting independent, non-grouped errors and its powerful encoding capacity of millions and billions of information combinations accompanied by its predetermined error correction capabilities (Table 1). For example, 128 different patterns can be generated with a 10 × 10 MNP and accurately decoded even when 15 error bits are present. Similarly, a 17 × 17 MNP can encode 137.4 billion different patterns and correct 31 error bits. This means billions of different patterns can be generated and used for encoding different medical information with a patch size of only 2 cm².

**Fig. 3: Deep learning-based networks allow parameter-free encoding and decoding of the OPMR.**

Table 1 Information capacity of 2D MNP array based on RM ECC

Full size table

Once the encoded binary string was generated, it was mapped into a 2D pattern with a fixed orientation (Fig. 3c). The generated patterns consist of 1-bits (ON) where microneedles are filled with NIR dye and 0-bits (OFF) where microneedles do not contain any fluorescent dye (Fig. 3d). Next, an encryption mask was added to the encoded pattern to ensure the privacy of personal medical data (Fig. 3e and Extended Data Fig. 4d). The encrypted pattern (Fig. 3f) was then encoded on a physical MNP. The encoded MNP was fabricated by selectively loading dye to only the ON-bit needles, yielding roughly 50% ON-bit and 50% OFF-bit needles (Fig. 3g).

Deep learning networks for spatially robust decoding

The spatial distribution of fluorescent dyes makes the OPMR system susceptible to bit distortions. The decoding phase was developed to compensate for these spatial variations between captured bit signals and to ensure spatial OPMR robustness. The decoding phase begins with the acquisition of raw images (Fig. 3h). Once acquired, each raw image was rectified to a square, binary format using a deep learning-based rectification network (Fig. 3i), which involves a U-Net-based³⁸ binarization network (Extended Data Fig. 5a) that converts the raw red–green–blue (RGB) image to a black-and-white binary image and a minimum area rectangle function (Extended Data Fig. 5b) that finds, crops and rotates the MNP region. After rectification (Fig. 3j), the image was fed into a deep learning-based recognition network (Fig. 3k), developed by training a convolutional neural network³⁹ (Extended Data Fig. 5c,d) with 650,000 synthetic images (Extended Data Fig. 5e–g). With these two deep learning steps, a raw image was successfully converted to a binary array (Fig. 3l). At this point, the binary array may have a corrupt pattern due to undetected or falsely detected signal bits. For accurate pattern decoding, the encryption mask was removed (Fig. 3m), and the deciphered pattern (Fig. 3n) was corrected with the RM ECC before being converted back to a binary string (Fig. 3o). The array was then, finally, translated and retrieved on a screen (Fig. 3p). The entire encoding-to-decoding workflow is completely automatic, requiring no user input or manual threshold manipulations due to the ‘end-to-end’ nature of this machine learning approach.

Long-term information data preservation in swine

Signal retention and pattern decodability were analysed longitudinally in a live pig model, as swine skin is known to be a close mimic of human skin³³. For signal retention (Fig. 3q), the number of NIR bits that were preserved in the skin was quantified over time (Extended Data Fig. 6a). Each of the ON and OFF bits was detected to find the percentage of ON bits out of the total transferred bits, and this value was output as the signal retention percentage. For pattern decodability (Fig. 3r), patterned MNPs were decoded and outputs were compared to the intended encoded information (Extended Data Fig. 6b). If the retrieved information accurately matched the original information, then the pattern was designated as successfully decoded; otherwise, it was designated as unsuccessful (Extended Data Fig. 6c).

MNPs were applied on the flank areas of Yorkshire pigs (Fig. 4a) and imaged weekly for three months. For signal retention, a total of 24 96-bit MNPs were applied on seven pigs. The dyes were invisible to the naked eye throughout the three-month monitoring period, resulting in completely indistinguishable patch application sites (Fig. 4b); however, the NIR signals remained visible during the entire monitoring period (Fig. 4c). For pattern decodability, a total of 21 patterned MNPs with four randomly selected 10 × 10 patterns (Extended Data Fig. 6d) were applied to three different pigs and imaged weekly for three months (Fig. 4d).

**Fig. 4: Long-term efficacy of OPMR in a swine model.**

The signal intensity of these applied patches decreased over time²⁸ (Fig. 2m), but with the scope of the camera and image acquisition (Extended Data Fig. 1a–c), NIR bits remained detectable, resulting in signal retentions of 98.69 ± 1.31% at 4 weeks, 98.35 ± 1.18% at 8 weeks and 98.44 ± 1.23% at 12 weeks (Fig. 4e). The number of error bits was well within the 15% correctable error bit threshold of the RM code of choice (Table 1). This completely automatic bit counting system processed 96-bit MNP images at an average speed of 0.043 second per image (n = 10,639 images). These machine learning-assisted results compared favourably to the adaptive threshold algorithm used previously in terms of both bit detection and precision (n = 20; Fig. 4f). For information preservation, all 21 patterned MNPs were decoded successfully across all three pigs over three months (Fig. 4g). This shows that our RM ECC successfully corrected the 1–2% of bit loss and retrieved accurate information over time. All 100% of the MNP footprints read out correct information despite evident spatial distortions caused by animal growth from 45.33 ± 10.78 kg to 94.33 ± 6.11 kg and epidermal cell turnovers during the three months (n = 3)^40,41 (Fig. 4h). This completely automatic pattern decoding system analysed 10 × 10, patterned MNP images at an average speed of 0.066 second per image (n = 2,385 images) on a laptop with an Intel Core i7 tenth generation processor, suggesting that automated decoding is unlikely to be a noticeable source of delay in an OPMR workflow.

Biocompatibility of the OPMR system

To understand the long-term biocompatibility of the OPMR, the cytotoxicity of the OPMR dye was first examined in vitro with no indication of cytotoxicity (Extended Data Fig. 7a–c). Next, local inflammatory and cellular responses in the area of MNP administration were examined. Upon MNP application, a mild erythema was initially observed and disappeared within 30 min (Fig. 2f). For histopathological evaluations, tissue sections applied with nothing and with MNPs containing polymer, blank PMMA microparticles and QD–PMMA microparticles were excised 3 days post-application. Cutaneous lesion scores for the QD–PMMA MNP group were comparable in severity to the other MNP control groups (Extended Data Fig. 7d), suggesting that the observed lesions were induced by trauma from needle penetration itself, irrespective of the PMMA or QD dye content⁴² (Fig. 4i). Skin sections retrieved 3, 30 and 70 days after QD–PMMA MNP application (Extended Data Fig. 7e,f) exhibited minimal to mild dermal inflammation at both days 30 and 70, a complete absence of hyperkeratosis at day 70 and no evidence of fibrosis at any examined time point (Fig. 4j). There were no clinically or statistically significant differences between cumulative histopathological scores of untreated and MNP-applied skin samples (Fig. 4k). Lastly, successful QD clearance from the skin tissue over time was also observed (Extended Data Fig. 7g).

Co-delivery of OPMR and mRNA vaccine for SARS-CoV-2

The OPMR MNP can co-deliver therapeutics to patients by adding a secondary cargo^43,44,45,46 (Extended Data Fig. 8). We demonstrated efficient and safe co-delivery of the OPMR and an mRNA vaccine encoding SARS-CoV-2 receptor-binding-domain (RBD) spike protein, encapsulated in LNPs in a rat model (Fig. 5a).

**Fig. 5: MNPs that co-deliver the OPMR and potent mRNA vaccine successfully record information and induce immunogenicity in rats.**

First, to evaluate the performance of the OPMR when co-delivered with mRNA encapsulated in LNPs, the pattern decodability of the OPMR was studied with and without mRNA-LNPs. For this test, 10 × 10 patterned MNPs (Extended Data Fig. 6d) with and without mRNA-LNPs (Extended Data Fig. 9a–d) were applied to Wistar rats and imaged for six months (Extended Data Fig. 9e). A 17 × 17 patterned MNP group was added to demonstrate the feasibility of recording billions of different patterns in the long term (Table 1). The footprints of both 10 × 10 (Fig. 5b) and 17 × 17 (Fig. 5c) MNP groups remained detectable and successfully decodable (Fig. 5d) during the entire six-month monitoring period (n = 5–6; Fig. 5e and Extended Data Fig. 9f). These 100% success rates indicate that a long-lasting OPMR is feasible with mRNA-LNP delivery.

Second, to study the efficacy of mRNA vaccine delivery with the OPMR, the integrity of LNPs was first characterized with and without the OPMR dye in vitro. When analysed with cryo-transmission electron microscopy (cryo-TEM; Fig. 5f) and dynamic light scattering (DLS; Fig. 5g and Extended Data Fig. 10a), the mRNA-LNPs remained stable and monodispersed with average sizes of 121.20 ± 1.56 nm and 120.20 ± 1.67 nm with and without OPMR dye, respectively. Both mRNA strands and LNPs remained intact with mRNA integrities of 93.60 ± 7.72% and 97.40 ± 2.59% (Fig. 5h), and mRNA encapsulation efficiencies of 87.50 ± 0.32% and 87.95 ± 1.28% (Fig. 5i), with and without OPMR dye, respectively, as well. Next, to evaluate the mRNA vaccine delivery with the OPMR in vivo, three groups were tested for immunogenic responses: (1) a control group with intramuscular (IM) injections, (2) an mRNA MNP group loaded with vaccine only and (3) an mRNA–OPMR MNP group co-loaded with vaccine and OPMR. For these groups, prime doses were applied on day 0, and booster doses followed 28 days after (Fig. 5a). All three groups exhibited comparable post-boost immunoglobulin G (IgG) titre levels (n = 6; Fig. 5j) as well as pseudovirus neutralizing antibody titre levels (n = 6; Fig. 5k), indicating that the OPMR MNPs successfully provided non-inferior protection against SARS-CoV-2 compared to the IM controls. This demonstrates that co-delivery of effective mRNA therapeutics is feasible with our OPMR MNP technology.

Finally, to evaluate the shelf life of the OPMR–mRNA MNPs, MNPs loaded with mRNA encoding firefly luciferase (FLuc) and OPMR dye were stored at room temperature for three months and applied to rats at different time points. When the bioluminescence of the FLuc expression was quantified with an in vivo imaging system at 6 h post-application (Fig. 5l), no substantial differences existed between fresh patches and those stored for one month or three months (n = 5; Fig. 5m), highlighting the possibility of storing, distributing and applying these patches on-demand for mRNA therapeutics delivery and recording.

Outlook

Here we developed a robust microneedle-based OPMR technology that can store information intradermally with excellent temporal and spatial robustness and an encoding capacity up to the billions. The demonstration of the co-delivery of the OPMR with reliable information retrieval and a potent mRNA vaccine exhibited in this work suggests a potential translation of the technology to clinical uses. In cases of emergency like in a pandemic or natural disaster, or at refugee or military camps, OPMR patches can be administered on-demand and can help healthcare workers make appropriate decisions on the follow-up dose administrations without patient confidentiality risks. To further strengthen the OPMR’s long-term reliability, different case scenarios (for example, pattern signal attenuation due to skin pigment, hair or pattern overlap; Extended Data Fig. 10b–d) for a prolonged period of time (for example, OPMR MNP stability for one year; Extended Data Fig. 10e–g) could be investigated in the future. Overall, this technology is readily applicable for any mRNA therapeutics, considering its compatibility with mRNA-LNPs and its large encoding capacity to complement the increasing number of mRNA therapeutics currently under development. As mRNA therapeutics aim to combat a wide range of preventable and incurable diseases, this OPMR technology presents an opportunity to bring healthcare equity one step closer to reality (Extended Data Fig. 10h).

Methods

QD encapsulation in PMMA microparticles

Commercial CuInS₂/ZnS QDs were purchased from Strem Chemicals. PMMA (molecular weight, ~120,000) was purchased from Sigma-Aldrich. QD encapsulation was performed following the solvent emulsification evaporation technique. Some 100 mg of QDs and 100 mg of PMMA were dissolved in 2 ml dichloromethane (DCM). The QD–PMMA solution was then added to 20 ml of cold 1% w/v polyvinyl alcohol (PVA) solution and emulsified at 10,000 rpm for 1 min (T 18 digital ULTRA-TURRAX homogenizer, IKA). The resulting emulsion was immediately poured into 30 ml of 1% w/v PVA solution and stirred at 250 rpm overnight to allow the evaporation of the DCM. Following this, the solution was poured into a 50 ml centrifuge tube and centrifuged at 2,400g to collect the microparticles. The supernatant was discarded and the microparticles were washed three times by adding sterile deionized water followed by centrifugation. Finally, the particles were resuspended in a small amount of water and filtered through a 75 μm filter to remove large aggregates. The QD-microparticles were then dried and kept in the dark, under vacuum until use.

Photoluminescence spectroscopy

Photoluminescence emission spectra were measured using a thermoelectrically cooled silicon camera (PIXIS100, Teledyne Princeton Instruments). Samples were prepared in quartz cuvettes by suspending QDs or QD–PMMA microparticles in cyclohexane. The samples were excited using a 532 nm laser (CPS532, Thorlabs). The emission was collected and focused using two silver-coated off-axis parabolic mirrors, and it was filtered through an 800 nm long-pass dielectric filter into a monochromator before imaging on the silicon camera. Photoluminescence quantum yield data were measured using a silicon photodiode (818-UV, Newport) coupled to a lock-in amplifier (SR830, Stanford Research), using chopped 405 nm laser excitation (LDM405, Thorlabs) and an optical integration sphere (RTC-060-SF, Labsphere)⁴⁷.

NIR fluorescence imaging

A universal serial bus (USB)-connected NIR fluorescence imaging system involving a custom light-emitting diode (LED) module that emits shorter-wavelength NIR light at 780 nm and a USB-connected camera module that captures the excited QD fluorescence image at longer-wavelength NIR light at >850 nm were used to image the QDs, with a photoluminescence intensity peak at 890–897 nm. An Android smartphone application ‘IR Record’ was custom developed to capture the OPMR NIR dye signal and can save images. The software is designed to take 30 consecutive images with six different exposure settings and five different gain settings. This bracket scanning method allows the capture of NIR signals with varying intensities over time. Among the 30 images, one image with the best reading results gets automatically chosen and processed. The total amount of required time to scan one OPMR patch on a patient is currently slightly over 2 min. However, further improvements can be made by pushing the scanning and recognition process to real time. To further advance the temporal aspect of the image-acquisition-to-image-recognition pipeline, (1) an ‘OPMR detection module’ that pinpoints the location of a relatively small patch, (2) an ‘auto-exposure module’ that acquires and stores one sufficient image for recognition instead of capturing a series of images with a wide window of exposure levels and (3) a ‘mobile-deployed decoding phase’ that processes everything on the fly can be added to our automatic system. With these implements, comprehensive scanning and recognition of numerous patterns will be executable in a more timely manner in real-world scenarios. In reality, we observed that the imaging environment (for example, ceiling lights in the surgical room of the pig facility, the height of the bed that the pigs were placed on and the distance between the bed and lights), the imaging personnel (for example, manual adjustment of the camera settings and distance away from the surface of the pig skin) and the pigs’ condition (for example, a scratch or fur on the pig skin) affect the quality of the images more than the actual quality of the NIR signals. These factors reflect real-world imaging scenarios more accurately as imaging will be performed by different healthcare professionals across different points of care.

Polydimethylsiloxane mould fabrication

Positive master moulds were designed using a computer-aided design (CAD) software (SolidWorks, Dassault Systèmes) and custom manufactured on a five-axis computer numerical control (CNC) milling machine using tool steel blanks. Master moulds with smaller needle spacings that were not suitable for CNC machining were three-dimensionally (3D) printed with HTL resin and ultra-high-resolution 3D printers (Boston Micro Fabrication). These master moulds were used to generate negative moulds made of polydimethylsiloxane (PDMS; Sylgard 184, Dow Corning). The PDMS base and crosslinking agent were mixed according to the manufacturer’s instructions, poured onto the positive master mould and cured overnight at 60 °C. To create additional MNP positives, UV-curable Norland Optical Adhesive 61 (Cranbury) was filled into the PDMS negative moulds using a centrifuge at 3,234g for 1 min, placed in a UV-curing oven at room temperature for 20 min and manually removed.

The mRNA-LNP solution synthesis

LNPs were synthesized to encapsulate mRNA^48,49,50. Pseudouridine-modified mRNA encoding FLuc (TriLink BioTechnologies) and pseudouridine-modified mRNA encoding the SARS-CoV-2 spike protein (National Institutes of Health code/strain) with furin cleavage site deletion, two proline mutations and a trimerization foldon for stability (ACROBiosystems) were purchased. For the LNP synthesis, Lipid 5 ionizable lipid (heptadecan-9-yl 8-((2-hydroxyethyl)(8-(nonyloxy)-8-oxooctyl)amino)octanoate; Organix), 1,2-dioleoyl-sn-glycero-3-phosphoethanolamine (DOPE; Avanti), cholesterol (Sigma-Aldrich) and 1,2-dimyristoyl-sn-glycero-3-phosphoethanolamine-N-[methoxy-(polyethyleneglycol)-2000] (ammonium salt; C14-PEG2000; Avanti) were dissolved in ethanol at a molar ratio of 38.4:12.3:47.4:1.9, respectively. To prepare the LNPs, the ethanoic solution was rapidly added to and mixed with an mRNA solution buffered with citrate at pH 3 at volume ratio 3:1 (aqueous/ethanol). The ionizable lipid to mRNA weight ratio was set to 5, and the final mRNA concentration was 0.135 mg ml^–1. All nucleic acids were stored at –80 °C and were allowed to thaw on ice prior to use. The LNPs were then dialysed for at least 2 h in phosphate buffered saline (PBS) at 4 °C in a 20,000 molecular weight cut-off cassette. For mRNA-LNP solution synthesis for MNP fabrications, LNPs were further dialysed in deionized water for an additional 2 h at 4 °C, and then concentrated to 320 µg ml^–1 on an Amicon filter by centrifuging at 3,000g. Finally, the mixture was diluted to a polymer solution made of PVA (Mowiol 4-88; molecular weight, 31,000; Sigma-Aldrich) and polyvinylpyrrolidone (PVP; molecular weight, 10,000; TCI) to form the vaccine solution.

OPMR microneedle fabrication

OPMR MNPs were fabricated using a centrifugation technique. To load the OPMR dye to the MNPs, 200 μl of an aqueous dispersion of QD–PMMA microparticles (3–10 mg ml^–1) was dispensed on top of a 10 × 10 negative PDMS mould and centrifuged at 2,400g for 3 min to concentrate the microparticles at the needle tips. To ensure an even loading of the dye across the MNP, the mould was rotated 180° before adding another 200 μl of the dye dispersion and was centrifuged for another 3 min. One more round of 200 μl of the dye dispersion loading and centrifugation was done without rotating the mould this time. For the microneedle body and backing, 150 μl of 30% w/w PVA/PVP (1:1) solution was added and centrifuged at 2,400g for 5 min. An additional 500 μl of the PVA/PVP solution was added and centrifuged in two steps with a 4 h time interval in between to ensure a flat surface of the backing after drying. After drying at room temperature for 4.5 days, a Delrin acetal plastic backing with adhesive tape (0.508 mm thickness; McMaster-Carr) was attached to the back of the MNP before removal from the mould. Once removed, it was further dried under vacuum (–60 MPa) for 48 h. As a result, each MNP ended up with a maximum of 0.8 mg QD–PMMA.

FLuc mRNA–OPMR MNPs were fabricated via a two-step loading using a vacuum-through technique. PDMS moulds were placed on a vacuum-through device to load cargos into the mould using negative pressure, leveraging the air permeable characteristic of PDMS. As the first step, 1 ml of an aqueous dispersion of QD–PMMA microparticles (1.5 mg ml^–1) with 0.5% w/w PVA/PVP was dispensed on the PDMS mould, and a vacuum (–85 MPa) was applied to the vacuum-through device overnight. As the second step, 200 μl of FLuc mRNA-LNPs mixed with PVA/PVP solution (1:320 mRNA to polymer ratio by weight⁴³) was dispensed and left to dry overnight under vacuum. As the final step, 80 μl of 20% w/w PVA/PVP solution was dispensed to form the MNP backing and was left to dry overnight under vacuum. Once dried, a Delrin backing was attached for patch removal, and the removed patches were further dried in a desiccator under house vacuum for 48 h and until use. The mRNA-only MNPs were fabricated the same way excluding the first step. For the shelf-life studies, FLuc mRNA–OPMR MNPs were fabricated and stored in a desiccator at room temperature for months. For each test time point, a fresh batch of mRNA–OPMR patches was fabricated as a positive control.

Patterned SARS-CoV-2 mRNA–OPMR MNPs were fabricated via the two-step loading using the vacuum-through technique. To make patterned MNPs, laser-cut mask tapes (468MP PEI Adhesive Transfer Tape Sheet, 3M) with patterns were placed on 20 × 20 PDMS moulds. Then, the PDMS moulds were placed on a vacuum-through device to load cargos into each mould using negative pressure, leveraging the air permeable characteristic of PDMS. As the first step, 1 ml of an aqueous dispersion of QD–PMMA microparticles (1.5 mg ml^–1) with 0.5% w/w PVA/PVP was dispensed onto the top of the mask on the PDMS mould, and a vacuum (–85 MPa) was applied overnight. Then, the mask was removed, and the mould surface was cleaned with an ethanol wipe. As the second step, 180 μl of vaccine solution made of SARS-CoV-2 RBD mRNA-LNPs (589 μg ml^–1 encapsulated mRNA) mixed with PVA/PVP solution at a 1:320 mRNA-to-polymer ratio by weight was dispensed and spread on top of the mould, covering approximately 100 needles in the centre, and was left under vacuum overnight. As the final step, a Delrin backing was attached for patch removal, and the removed patches were further dried in a desiccator under house vacuum for 48 h and until use. As a control group, SARS-CoV-2 RBD mRNA MNPs without the addition of OPMR dye were fabricated; this group was fabricated the same way excluding the first step. For 17 × 17 patterned patches, no mRNA was loaded. Instead, after the dye loading, 1 ml of 20% w/w PVA/PVP polymer solution was added and spread across the 17 × 17 array to form the needle body and backing. The process to deliver a sufficient dose of mRNA vaccine with the MNP is shown in a design relationship plot in our recent publication⁴³, created to meet dosing requirements in humans for common mRNA vaccines; it illustrates the relationship among single microneedle volume, microneedles per MNP and dose delivered in a single MNP. Using the needle volume and MNP loading efficiency from a study on anti-COVID RBD titre peaks, this model predicts the combination of volume and number of microneedles necessary to deliver the full Moderna (mRNA-1273) or Pfizer–BioNTech (BNT162b1) COVID-19 vaccine doses. The model predicts that 360 and 108 of the studied microneedles and formulations would deliver the full dose of the Moderna and Pfizer–BioNTech vaccines, respectively. MNPs containing a sufficient dose are less than 2 cm across.

Scanning electron microscopy analysis

SEM was used to image the QD–PMMA microparticles and the OPMR and OPMR–mRNA MNPs. Samples were initially coated with a thin layer of Au using a sputter coater (Desk V, Denton Vacuum) and then imaged using high-resolution SEM (Zeiss Crossbeam 540 Scanning Electron Microscope and Focused Ion Beam, Zeiss).

Ex vivo and in vivo MNP applications

All animal procedures were approved and performed under the guidelines of the Massachusetts Institute of Technology Committee on Animal Care. Three-month-old female Yorkshire pigs were provided by Cummings School of Veterinary Medicine. Six-week-old to eight-week-old female Wistar rats were purchased from Charles River. For the ex vivo MNP applications, hand application, a commercial spring applicator (Micropoint Biotechnologies) and custom-made applicators with a velocity upon impact ranging from 1,066–1,583 cm s^–1 and holding force within the range of 0.16–1.98 MPa were tested on excised flank and hip skins of pig cadavers. For ex vivo human skin tests, donated cadaveric skin tissue samples (National Disease Research Interchange, Philadelphia, PA) were used. For in vivo MNP applications, a custom-made applicator with a velocity on impact of 1,407 cm s^–1 and holding force of 1.1 MPa was used to apply patches on the flank and hip areas of Yorkshire pigs and on the back area of Wistar rats. Patches were applied for 5–10 min for pigs and 10–20 min for rats.

Dye deposition depth assessment

To assess the dye deposition depth in the skin, skin tissues were excised after MNP applications, fixed in 10% formalin buffer for 48 h, transferred to 70% ethanol and then embedded in paraffin wax. Samples were sectioned at 5 μm width every 20 μm to retrieve 30 slices for one entire microneedle array for cross-sectional evaluations of the maximum needle penetration and dye deposition depths. The samples were stained with hematoxylin and eosin and analysed with Aperio software (Leica Biosystems). Furthermore, parts of the skin tissue were frozen and fixed in Optimal Cutting Temperature compound for cross-sectional imaging to detect the presence of the NIR bits in the dermis.

Needle dissolution, bit transfer, signal retention and signal intensity analyses

To quantify microneedle dissolution as a function of different MNP architecture variables, MNPs were imaged before and after application using a Leica DFC450 optical microscope. The patches were placed in a transverse manner for imaging using LAS v.4.7 software. Microneedle length was calculated using ImageJ (National Institutes of Health) for >10 microneedles per MNP for 3–5 MNPs per group. For bit transfer and signal retention analyses, ImageJ and the adaptive threshold were used to count the number of bits from captured NIR images at different time points; these measurements were performed on five patches per group. For the signal intensity analysis, ImageJ and the adaptive threshold were used to evaluate the maximum pixel value in each NIR bit in captured images at different time points; these measurements were performed on 50–96 bits per patch for three patches per group. To evaluate the signal intensities between different time points, a consistent gain and exposure combination was used for fair comparisons.

RM ECC for 2D binary array generation

Information bits of interest were encoded with redundancy using a RM ECC. The Python package ‘reedmuller’ (https://pypi.org/project/reedmuller/) was used for RM error correction encoding and decoding. The encoded binary string was then mapped to a 2D binary array after referring to a 2D template for orientation. The RM code with parameters r and m, denoted the RM(r, m) code, encodes (k={sum }_{i=0}^{r}left(begin{array}{c}m\ iend{array}right)) information bits with 2^m string bits and is capable of correcting c = (2^m–r–1 – 1) independent error bits. There is a trade-off between the number of information bits k and the maximum correctable error bits c, depending on the order r. For a 10 × 10 MNP with a RM(1, 6) code, the number of available string bits is ({2}^{leftlfloor {log }_{2}(10times 10)rightrfloor }=64), where ⌊⌋ is the floor function, which results in encoding 7 information bits into a 64-bit string. The rest, 36 bits, were reserved for orientation correction. Nine orientation bits (3 × 3 size) were allocated to four corners of the 2D array (three corners ON and the bottom right corner OFF) for patch orientation and registration of rotation angles. Encoded binary string bits were arranged sequentially from top left to bottom right (omitting orientation bits at the four corners) in a raster scan manner. For a 12 × 12 MNP with r = 2, the RM(2, 7) code can encode 526.8 million different information combinations and guarantee to correct up to 15 error bits. For a 17 × 17 MNP with r = 2, the RM(2, 8) code can encode 137.4 billion different information combinations and guarantee to correct 31 error bits. This means that the desired error correction capability can be predetermined and one can select a RM(r, m) option to fit the use case. With a RM ECC, the proposed OPMR technology can generate 137.4 billion different patterns with 1.7 cm × 1.7 cm MNPs, which can accommodate the large number of varieties of vaccination information, the eight billion people in the human population on Earth or the fast-growing number of therapeutics that are available or currently under development.

Encryption mask for patient data privacy

The OPMR system provides a high level of medical privacy, achieved through the strategic implementation of advanced security-preserving technologies across three layers. First, we use NIR fluorescent QDs that require a specific excitation wavelength (short-wavelength NIR at 780 nm) and a high-pass or band-pass filter mounted on the phone to receive only the emitted light (at the peak of 890–897 nm). Second, similar to password-protected hardware or software, we have a specific encryption procedure implemented in our OPMR system. Adding a known and fixed encryption pattern ensures the privacy of the personal medical data in the OPMR system. To decipher the encoded information, one must know not only the forward encoding method, but also the decipher key (that is, the fixed encryption mask). An encryption mask, composed of roughly half ON bits and half OFF bits, was added to the initially generated 2D pattern in a pixel-wise manner, applying a logical exclusive operator or an XOR operator on the initial encoded pattern. If an ON bit existed on the same pixel coordinate of both the encoded pattern and mask pattern, the pixel value was flipped. This step made the number of ON and OFF pixels even on the MNP, which made the recognition system robust to any pattern during the decoding step, because a RM code is highly structured and can have limited pixels in certain region or rows/columns. After randomly flipping pixels on the raw encoded pattern, the encrypted pattern will consist of half ON and half OFF pixels on average, which makes the recognition system robust to any patterns during the decoding step.

Synthetic image generation for deep learning network training

Simulated fluorescence images were constructed for training deep learning networks. The synthetic training set included a variety of patterns, NIR bit spacings, locations, intensities, background noise levels and contrasts, MNP scales, rotations (from –5° to 5°), distortions, image qualities, defocusing levels, motion blurs, brightnesses, exposures, gains and more to consider potential image formation variations at three levels: the formation of a MNP, the formation of an image and the acquisition of a fluorescence image. To capture the temporal variations, the percentage of ON pixels was varied by 100%, 75%, 50% and 25%, for the 5%, 15%, 60% and 20% portions of the image dataset, respectively. A deep learning-assisted rectification network was applied to these synthetic models to output paired results. Detailed parameter settings and code for synthetic image generation can be found at https://github.com/liuyang12/ecc-microneedle. These paired results were input (90% for training and 10% for validation) to a convolutional-neural-network-based recognition network for it to learn the mapping of the input images to binary arrays. We used paired data to train models for 10 × 10 and 17 × 17 MNPs separately, each with 650,000 images (90% for training and 10% for validation). This large number in the synthetic training set improved the robustness of the recognition system overall. For all network training and analysis, Microsoft Excel v.2021, GraphPad Prism v.10, FIJI v.2017, Python v.3.8, PyTorch v.1.11 and MATLAB v.R2023b were used.

Deep learning network for image binarization

The image binarization network was directly adapted from an off-the-shelf convolution network, U-Net, which was originally proposed for biomedical image segmentation. The binarization network took a single-channel (greyscale) 256 × 256 image as an input and output a two-channel 256 × 256 mask for final binarization. We used the cross-entropy loss function as the training criterion. During training, we used the same data generation process as the recognition network. This was an image-based ConvNet, which is light and accurate and easier to train with a given amount of training examples. To address binarization issues caused by impulse noise, we used 250,000 additional images with impulse noise augmentation (that is, salt and pepper noise, random erasing small rectangles and random Gaussian blur).

Minimum area rectangle

As the second rectification step, a rectangle with the minimum area that covers all the white bits of the 2D array region was generated using an off-the-shelf Python implementation by OpenCV, ‘cv.minAreaRect()’. The selected region was then rotated, cropped and resized to a target size. The final crop size was 35% larger than the size of the minimum area rectangle while preserving the centre of the rectangle as a reference point. This rectification step was essential for efficient network training of the recognition model because keeping 2D arrays centred and normalized to the same scale reduces the amount of spatial variety (that is, data augmentation for training).

Convolutional neural network for image recognition

We used a convolutional-neural-network-based network structure for the recognition model. We used the same network whether the MNP was 10 × 10 or 12 × 12. The input image sizes were 120 × 120 and 136 × 136, respectively, and these two models were trained separately using training samples with corresponding sizes. The input single-channel image was convolved by a 3 × 3 kernel with zero padding (adding zeros to the boundaries to keep the same image size) and then down-sampled to half the size. The same 3 × 3 convolution and down-sampling were repeated two times. The second convolution layer used no padding to obtain the target binary array size. After three convolution plus down-sampling layers, the 128-channel tensor was convolved by a 5 × 5 kernel without padding and finally convolved by a 1 × 1 kernel to get a two-channel mask for the final binary array. We used the cross-entropy loss function as the training criterion. BinarizationNet required the MNP size (N × N) as the input, while it was independent of the MNP size. The current OPMR system uses a laptop to run Python-based image processing and pattern recognition codes. In the future, this fully parameter-free, end-to-end structure could be converted to a Java-based smartphone application and combined with our image acquisition device to make it a stand-alone mobile system. This plug-and-play module could pave the way for easy deployment of the OPMR. These advancements would bring the OPMR technology one step closer to being clinically translational and readily applicable in the field. For all network training and analysis, Microsoft Excel v.2021, GraphPad Prism v.10, FIJI v.2017, Python v.3.8, PyTorch v.1.11 and MATLAB v.R2023b were used.

Histological evaluation of OPMR MNP applications in swine

Pig skin tissues were excised after various time points post OPMR MNP application for histological semi-quantitative evaluations of the effect of the dye in the skin over time. The skin sections were evaluated by a board-certified veterinary pathologist (S.E.C.). Sections were examined and photographed using an Olympus BX45 microscope attached to a DP26 digital camera (Olympus). Skin lesions were graded for epidermal hyperplasia, hyperkeratosis, neutrophilic and eosinophilic inflammation, mononuclear leucocytic inflammation and dermal fibrosis. Neutrophilic/eosinophilic inflammation, mononuclear inflammation and hyperkeratotic lesions were graded with a numerical score from 0 to 4, in which 0 = normal, 1 = minimal, 2 = mild, 3 = moderate and 4 = severe. Epidermal hyperplasia and dermal fibrosis were graded from 0 to 3, in which 0 = normal, 1 = mild, 2 = moderate and 3 = severe. The scores for each parameter were averaged from 6–12 skin sections from 3–6 patches per group from 2–3 pigs. Specific histopathology scores between experimental groups or between time points were compared by a Mann Whitney U-test.

Additionally, quantification of cleaved caspase-3 (CC3) staining of pig tissue with no treatment, and with MNP applications loaded with just PVA/PVP polymer, blank PMMA microparticles or QD–PMMA microparticles were applied three days prior to skin excision. To study if the OPMR system activates cell apoptotic mechanisms, tissue samples where the OPMR dye was deposited were stained with CC3, which is a marker used to detect changes in cellular morphology (for example, shrinkage and degeneration, nuclear condensation and fragmentation) and which is particularly useful to detect apoptotic cells. The 2D image analysis of ten tissue cross-sections from each group tested and stained with CC3 was performed using the open-source software QuPath and ImageJ. An area approximately 2 mm in depth and 8 mm in length across the epidermis surface was manually defined for all images. Preprocessing steps were applied on each image to prepare them for subsequent analysis using an image analysis thresholder. The thresholder was created to isolate tissue samples from the image background and designate these sections as regions of interest in the software. First a colour transform on the image was applied, which provided a clear and binary way to contrast between positive (brown) and negative (purple) stained areas in the tissue. Then, the regions of interest were exported to ImageJ. In the exported binary image, the threshold of intensity indicating the positive staining was set. This threshold was set by manually checking, with trial and error, which threshold resulted in a more accurate depiction of positive versus negative staining. Using the threshold function of ImageJ, we were able to find the percentage of tissue area that clears that threshold, which corresponds to CC3-positive regions. Quantitative analysis of CC3 staining showed no differences in the apoptotic cell percentage between the control and experimental groups, indicating there was no signal of immunoreactivity in the skin sections.

OPMR dye cytotoxicity analysis

For the OPMR dye (QD nanoparticles encapsulated in PMMA) cytotoxicity analysis⁵¹, HeLa cells were cultured in high glucose Dulbecco’s Modified Eagles Medium with phenol red (DMEM, Invitrogen) supplemented with 10% fetal bovine serum (Invitrogen) and 1% antibiotic (Invitrogen). Some 5,000 cells were seeded in a 96-well plate in full growth medium. Twenty hours after seeding, the media were replaced with fresh media including QD–PMMA microparticles at different concentrations and the cells were incubated for 20 h. Then the media were removed, cells were washed once with PBS buffer, MTS (Abcam) was added at 20% v/v in DMEM, cells were incubated for 4 h and absorbance was measured at 490 nm.

Additionally, human dermal fibroblasts were also cultured in fibroblast growth media (Invitrogen) supplemented with 1% penicillin/streptomycin (Invitrogen). Cells were seeded at a density of 5,000 cells per well in a 96-well plate containing full growth medium. Twenty hours post-seeding, the media were replaced with fresh media containing varying concentrations of QD–PMMA microparticles, alongside a fresh media control, and the cells were further incubated for 20 h. To assess cell viability, the live–dead assay and Cell Counting Kit-8 (CCK-8) were conducted. For the live–dead assay, cell media were aspirated, and the cells were gently washed once with PBS buffer. Cell viability was evaluated using the LIVE/DEAD Viability/Cytotoxicity Kit (Invitrogen, L3224) according to the manufacturer’s protocol and imaged under a DeltaVision Ultra microscope. The ratio of live cells to total cells was quantified as the ratio of viable cells to total cells using ImageJ. For the CCK-8 assay (Sigma-Aldrich), following media removal and PBS washing, cells were incubated with CCK-8 solution for 4 h and absorbance was measured at 450 nm. The viable cell number from each experimental group was normalized to the untreated control.

The mRNA-LNP concentration and encapsulation efficiency analyses

The mRNA concentration and encapsulation efficiency in the LNPs was quantified using a Quant-iT RiboGreen assay (Thermo Fisher) and a modified procedure described elsewhere⁴³. The encapsulated mRNA in LNPs were evaluated by quantifying the difference of mRNA concentrations in ×1 Tris–EDTA buffer and in 4% Triton X-100 buffer. The concentration of total mRNA was quantified by diluting mRNA-LNPs in Triton X-100 buffer. To quantify mRNA loading in MNPs by mass, microneedles were cut and dissolved in Tris–EDTA and Triton X-100. Subtracting the unencapsulated mRNA from the total mRNA yields the mRNA encapsulation efficiency.

The mRNA and LNP quality assessments

The mRNA-LNPs were qualitatively assessed using cryo-TEM with and without QD–PMMA microparticles. Samples (0.1–0.5 mg ml^–1) were added onto the carbon-coated copper TEM grids and the excess solution was blotted. Next, samples were plunge-frozen using a 930 Gatan Cryo-Plunge3 (Gatan). All the samples were imaged using a JEOL 2100 field-emission gun microscope (JEOL) at 200 kV acceleration voltage. Hydrodynamic size and polydispersity of mRNA-LNPs were analysed using DLS on a Zetasizer Nano-NS (Malvern Instruments) with and without QD–PMMA microparticles. Some 5 µl of the samples was diluted in 995 µl of UltraPure water in polystyrene cuvettes for size measurements. Three technical replicates were performed for each sample. The mRNA integrities in LNPs were assessed using a FEMTO pulse RNA fragment analyser (Agilent Technologies)⁵² with and without QD–PMMA microparticles. Some 2 μl of the samples with 0.5 pg μl^–1 dilution in UltraPure water was loaded per lane. Three technical replicates were performed for each sample.

SARS-CoV-2 vaccination with mRNA–OPMR MNPs in rats

Six-week-old to eight-week-old female Wistar rats were used for immunogenicity tests from vaccination via IM injection, mRNA MNP administration and mRNA–OPMR MNP administration. Each MNP contained 10 µg of encapsulated mRNA encoding SARS-CoV-2 S-protein RBD, and as a positive control, a matching dose of fresh mRNA-LNPs suspended in PBS was administered intramuscularly (n = 6 per group). MNPs were applied to the back area using an applicator for 20 min, and IM injections were administered to the thigh muscle. All animals received a booster dose via the same method (that is, IM injection, mRNA MNP or mRNA–OPMR MNP) at 28 days after the prime dose. Rats were bled at 3 and 7 weeks after the prime dose, and enzyme-linked immunosorbent assay (ELISA) was performed to evaluate anti-RBD binding titres.

SARS-CoV-2 anti-RBD binding titres

Anti-RBD binding titres in rats were analysed using SARS-CoV-2 S-protein RBD ELISA detection. Recombinant SARS-CoV-2 S-protein RBD (1 μg ml^–1, 100 μl per well, overnight at 4 °C; ACROBiosystems, SPN-C52H9) was used to capture anti-RBD IgG titres in rat serum (serial dilutions in PBS, 2 h at 37 °C). Goat polyclonal antibody (pAb) to rat IgG horseradish peroxidase conjugates (Abcam, ab112767) were used as secondary antibodies (1:10,000 dilution in blocking buffer, 100 μl per well, 1 h at 37 °C), and 3,5,3′,5′-tetramethylbenzidine (TMB) was used as a substrate (100 μl, 5 min incubation before addition of 100 μl of 3 N H₂SO₄). End-point titres were calculated as the dilution that emitted an optical density exceeding ×3 of the background produced by serum from naive mice.

Pseudovirus-based neutralization assay

The SARS-CoV-2 pseudoviruses WA1/2020 strain (Wuhan/WIV04/2019, Global Initiative on Sharing All Influenza Data (GISAID) accession no. EPI_ISL_402124), expressing a luciferase reporter gene, were generated. In brief, the packaging plasmid psPAX2 (AIDS Resource and Reagent Program), luciferase reporter plasmid pLenti-CMV Puro-Luc (Addgene) and spike protein expressing pcDNA3.1 SARS-CoV-2 SΔCT of variants were co-transfected into HEK293T cells (ATCC, mycoplasma tested) using Lipofectamine 2000 (Thermo Fisher). The supernatants containing the pseudotype viruses were collected 48 h after transfection, and then were purified by centrifugation and filtration with a 0.45 µm filter. To determine the neutralization activity of the plasma or serum samples from participants, HEK293T-hACE2 cells were seeded in 96-well tissue culture plates at a density of 1.75 × 10⁴ cells per well overnight. Threefold serial dilutions of heat-inactivated serum or plasma samples were prepared and mixed with 50 µl pseudovirus. The mixture was incubated at 37 °C for 1 h before being added to HEK293T-hACE2 cells. Forty-eight hours after infection, cells were lysed in Steady-Glo Luciferase Assay (Promega) according to the manufacturer’s instructions. SARS-CoV-2 neutralization titres were defined as the sample dilution at which a 50% reduction in relative light unit was observed relative to the average of the virus control wells. The results were analysed using ordinary two-way ANOVA (Sidak’s multiple comparisons test).

Firefly luciferase mRNA expression in rats

Six-week-old to eight-week-old female Wistar rats were used to test the delivery of mRNA encoding FLuc when administered with OPMR MNPs. MNPs loaded with 5 μg FLuc mRNA encapsulated in LNPs with or without OPMR dye in 17 × 17 arrays were fabricated and applied to the back area using a custom applicator with a velocity of 1,407 cm s^–1 and holding force of 1.1 MPa for 10 min. Six hours after application, rats were imaged for bioluminescence of luciferase expression using an in vivo imaging system that was also a kinetic imaging system (PerkinElmer). Fifteen minutes prior to imaging, rats were injected with IVISbrite d-luciferin potassium salt XenoLight (PerkinElmer) intraperitoneally at 150 mg kg^–1. Luminescence was quantified using LivingImage software (PerkinElmer).

Statistical analysis

All in vitro and ex vivo experiments were performed in experimental triplicate or quintuplicate unless noted otherwise. All in vivo experiments were performed with five or six experimental replicates unless noted otherwise. Statistics analyses were performed using GraphPad Prism software using a two-tailed Student’s t-test for pairwise comparisons (non-statistical significance, P > 0.05). For multiple comparisons, one-way ANOVA was used unless noted otherwise. In all figures, data are presented as mean values, and ±s.d. is used for error bars.

Ethics and inclusion declarations

Massachusetts Institute of Technology is dedicated to offering a safe, respectful, friendly and collegial environment for the benefit of everyone who attends, and for the advancement of the interests that bring us together. All animal procedures were approved and performed under the guidelines of the Massachusetts Institute of Technology Committee on Animal Care. Three-month-old female Yorkshire pigs were provided by Cummings School of Veterinary Medicine (protocol 0919-058-22). Six-week-old female Wistar rats were purchased from Charles River (protocol 0916-057-20).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.